ai-rank
HOW IT WORKS

How we make sure the report can be trusted.

An AI assistant can be wrong: it invents, confuses, forgets. That is why every name, fact and source we show you has a proof, and a person rereads everything before it reaches you. Here we explain, without jargon, what we check and how.

  • Every data point has proof
  • A person rereads everything
  • Delivered within 48 hours

Why it is not the same as asking an AI assistant yourself.

You can open an assistant and ask “can you recommend an accountant in Milan?”. But one answer cannot tell you whether you appear: it changes with every request, you cannot tell if it is true, and it does not tell you what to do.

If you ask
1

answer, from one assistant, to a question written on the fly

The ai-rank report
30

answers, from 3 advanced models, to 10 questions written for you

Asking an assistantThe ai-rank report
The questions

One or two, written as they come, often with your name in them.

10 questions as a new customer would ask them, in your market's language, approved by a person.

The answers

One answer, which changes every time you ask. A single try tells you nothing.

30 answers from 3 advanced models with web search (more powerful than the ones in the free test), all kept and read one by one.

The result

An impression: “it named me”, “it did not name me”.

A score from fixed rules, calculated only on the questions that do not name your business: what a new customer would see.

What is true

You do not know whether what it says is right or invented.

Every data point has a verified quotation. What is said about you is compared with your website and errors are flagged.

Competitors

A few names, without knowing how much they weigh.

Who is named instead of you, in how many answers and by which assistants, on a positioning matrix.

Sources

You do not know where the answer comes from.

The cited pages, grouped, with your site separated from profiles and directories, and the sentence each piece of advice comes from.

What to do next

No guidance, or generic advice.

10 actions in priority order, reviewed by a person, each with the data it comes from.

Time

Hours of trying and comparing by hand, with no method.

Delivered within 48 hours of receiving your details.

From your website to the report, in seven steps.

No step is left to automation alone: there is a check between one phase and the next.

  1. 1

    We read your website

    We start from what you say about yourself: services, city, the clients you serve. So the questions are about your business, not a generic one.

  2. 2

    We write 10 questions

    They are the questions a real customer would ask, in your market's language. A second check rereads them to remove repetitions and invented details.

  3. 3

    A person approves them

    One of our reviewers reads the questions and the list of competitors before the answers are collected. If a question makes no sense for your business, they change it.

  4. 4

    We ask 3 advanced AI models

    Each question goes to three advanced models with web search, more powerful than the ones in the free test: 30 answers, all kept in full. If any is missing, we ask again.

  5. 5

    We read every answer

    For each one we record who is named, who is recommended, which sources are cited and what is said about your business.

  6. 6

    We verify

    A name, a fact or a source enters the report only if it is really written in the answer. What is said about you is compared with your website.

  7. 7

    A person rereads and delivers

    The reviewer checks competitors, sources, doubtful information and actions, and can have the analysis redone. Only then does the report reach you, within 48 hours of receiving your details.

What we check, in detail.

Four things that can go wrong in an AI report, and what we do about each.

COMPETITORS

Are the competitors really competitors?

The question we ask ourselves

Are the names AI recommends instead of you real businesses, in the same trade and your area?

How we check it
  • ✓We only show names written in the assistants' answers. We do not add or remove any by hand.
  • ✓We merge variants of the same name (“Bridgeway” and “Bridgeway Accountants Manchester”), so one business is not counted twice.
  • ✓For each name we count in how many questions and with how many assistants it appears: a name cited once weighs less than one cited in twenty answers.
  • ✓Competitors you list yourself are checked before the answers are collected.
  • ✓The reviewer reads the final list and, if they find portals, directories or businesses in another city, writes it in the report as a recognisable note and takes it into account in the actions.
If something does not add up

The numbers stay as measured: we do not touch them. But the conclusions do not rely on a name that is not a real competitor, and the reviewer's note tells you so.

SOURCES

Are the sources real and relevant?

The question we ask ourselves

Do the pages the assistants cite really exist, and are they the ones that matter for your business?

How we check it
  • ✓Sources are the addresses the assistant opened or cited with web search, not names written from memory.
  • ✓We group them by site and count how often they appear, separating your own site from third parties: Google profiles, professional registers, directories, review sites.
  • ✓When a suggestion comes from a cited page, the address stays linked to the sentence it comes from, so you can check it.
  • ✓The reviewer looks at whether the heaviest sources are relevant and whether you are present on them (complete profile, reviews).
If something does not add up

A source that is not relevant is marked as such by the reviewer, who can ask that the actions not be based on it.

MODEL ERRORS

Did the AI invent something? Did we?

The question we ask ourselves

Is what the assistants say about you correct? And is what we read in the answers really written there?

How we check it
  • ✓Every name, fact or hint we extract from an answer must come with a quotation that really appears in the text. A program checks it word by word: whatever cannot be found is discarded.
  • ✓What an assistant says about you (address, services, prices, experience) is compared with your website. If it contradicts it, it goes under “Information to verify”, with the exact sentence and the model that said it.
  • ✓The reviewer rereads every flagged item: sometimes the model is wrong, sometimes your website or profile is out of date, and the fix is different.
  • ✓The score is not calculated by a model: a program calculates it with fixed rules, the same for everyone.
If something does not add up

An answer that fails the checks does not count, and the report says how many valid answers it is based on (for example “28 of 30”). If too many answers are missing, we ask again instead of delivering an incomplete report.

ADVICE

Does the advice fit your business?

The question we ask ourselves

Are the 10 actions made for you, or the usual generic list?

How we check it
  • ✓The actions come from the facts found in your answers: where you do not appear, who wins, which sources weigh, which errors exist. A program discards any action that cites data that does not exist.
  • ✓Each action explains why: it points to the data it comes from.
  • ✓The priority order is calculated by a program from impact and effort, not chosen by a model.
  • ✓The reviewer reads every action: if it is too generic or does not fit your case, they have it rewritten, saying what to stress or leave out and adding information only they know, marked as such.
If something does not add up

An action that does not rest on real data does not reach the report.

An example of review.

Three typical cases of what the person who rereads the report does.

LESS RELEVANT SOURCE

A generic portal weighs more than it should

What shows up

An assistant often cites a classifieds portal as a source for “accountant in Milan”.

What the reviewer does

The reviewer flags it: it is a generic listing, not a source to work on. The actions point instead to the Google profile and the professional register, where rivals are stronger.

ERROR TO VERIFY

Old address: is the model wrong, or the website?

What shows up

An assistant gives “Via Dante 12” as the address, but the website says “Corso Venezia 8”.

What the reviewer does

The reviewer checks the Google profile and directories: the old address is still there. It is not the model's fault. The action becomes: update the address everywhere, in the same format.

ACTION TOO GENERIC

“Write more content” is not enough

What shows up

The draft of an action just says to publish more content on the website.

What the reviewer does

The reviewer has it rewritten: a dedicated page on “VAT numbers”, with indicative prices and the questions clients really ask, because that is the topic on which two rivals are cited 9 and 12 times.

Illustrative examples with made-up data, as in the demo report.

Who checks, and what they can do.

The review is done by a person on the ai-rank team, not by a model. They read the questions before the answers are collected and the report before delivery.

ai-rank is a service of Star Digital LLC. To know who we are, see the page About us.

They can

  • ✓change or discard questions
  • ✓write notes and information only they know, recognisable in the report
  • ✓have the analysis redone, saying what to stress
  • ✓hold back delivery until something adds up

They cannot

  • ✕touch up the score
  • ✕add or remove competitors from the measured data
  • ✕change what the assistants answered
  • ✕let through data with no proof in the answers

What we cannot promise.

  • The report is a snapshot of the day it was made: assistants' answers change from one day to the next and from one request to another.
  • It is not a ranking and it is not about Google.
  • We do not promise customers or positions: we tell you what to change and why.
  • If the data is not enough for a conclusion, we say so instead of guessing.

Want to see what it looks like?

See the demo reportTry the free test →